The field of artificial intelligence (AI) has undergone several important conceptual changes over the years. The classical approach (also known as GOFAI= good old fashioned AI) assumed that the mind can be modeled by the software of a computer and focused on abstract symbol manipulation that is independent of the specifics of a particular hardware. Connectionist models attempted to account for the 'hardware' of biological brains and simulated the massively parallel pattern processing that occurs in neurons.  Finally, the embodied approach accounts for the interactions of intelligent agents with their environment and acknowledges that intelligence requires a body.  Clearly, How the body shapes the way we think defends the third approach.  Authors Rolf Pfeifer and Josh Bongard set themselves a twofold goal: (ii) to show how having a body affects intelligence and (ii) to explore the conceptual implications of embodiment (p. xviii).  To achieve this goal they employ a synthetic methodology (understanding by building) and develop a theory of intelligence that consists of a set of design principles for intelligent agents who have physical bodies and exhibit a broad range of behaviors in their interactions with the environment.  The twelve chapters of the book are grouped in four thematic parts: Part I (Intelligence, Artificial Intelligence, Embodiment, and What the Book is About) introduces the reader to the basic concepts; Part II (Toward a Theory of Intelligence) develops the authors' theory of intelligence and embodiment, Part III (Application and Case Studies) applies the concepts and design principles developed in Part II to concrete problems such as ubiquitous computing, business and management, and the psychology of human memory, and Part IV (Principles and Insights) summarizes the major points and provides a review of the design principles.

Before developing their own theory of intelligence Pfeifer and Bongard introduce the reader to the mind-body problem (focus box on page 6) and provide a concise overview (from Descartes to Libet and Chalmers) of attempts to solve this problem.  Further, they explain traditional and contemporary concepts of 'thinking', 'consciousness', 'cognition', 'agency' and how these concepts relate to 'intelligence'.  Probably the most significant aspect of the theory of intelligence developed by the authors is its complete breach with any form of dualism.  Pfeifer and Bongard explain that Cartesian substance dualism (the belief that the body is controlled by an immaterial mind) has been subjected to devastating criticism for decades but that the belief that the brain controls the body on the conscious and subconscious level (e.g., thought, action, digestion, heartbeat) still appears plausible to many AI researchers.  By contrast, for Pfeifer and Bongard brain and body are equally involved in intelligent behavior because both are required to allow agents to "comply with the physical and social rules of their environment and exploit those rules to produce diverse behavior" (p.16).  The authors emphasize that on the one hand, even the most complex cognition tasks are "highly constrained by our bodies" (p.2).  On the other hand, embodiment simplifies the information processing demands of many tasks (e.g., sense perception, neural control of movement) that are involved in cognition because the morphological and material properties (e.g., stiffness and deformability of fingertips, legs, and feet; springy material of the muscle tendon system; positioning of eyes and ears etc.) of the agent and his interactions with the physical environment can be exploited.  Thus, embodiment is a prerequisite for any kind of intelligence (p.19) and Pfeifer and Bongard's theory of intelligence focuses on the interactions between body and brain.  They point out that intelligent artifacts have to function in the real world and that the design of such artifacts requires a synthetic methodology ("understanding by building", p. 21).  Departing entirely from computational theories of intelligence Pfeifer and Bongard define their core approach to intelligence as "a set of design principles that on the one hand represent fundamental ingredients for a general theory of intelligence and on the other provide powerful engineering heuristics for the design of intelligent artifacts" (p. 65).  The design principles can be used in an analytic way for understanding intelligence and in a synthetic way for building intelligent artifacts.

Pfeifer and Bongard introduce design principles for (intelligent) agents, development, evolution, and collective systems.  Naturally, agent design takes centre stage and many of the design principles employed here are informed by the study of living organisms and already existing artifacts (e.g., parallel, loosely coupled processes, sensory motor coordination, cheap design, redundancy).  Development of individual agents over time is also well studied for biological organisms and the relevant design principles (e.g., integration of time scales, incremental nature of development, discovery, social interaction) are increasingly taken into account for robot design.  Much more work remains to be done regarding the principles of evolution of intelligent artifacts (e.g., population, cumulative selection and self organization, scalable complexity).  The main reason for this is that the study of evolution requires long time scales and large populations of interacting agents.  While computer simulations can overcome the time-scale problem they do not allow for real world interactions and will always be limited by the input criteria selected by the designer of the simulation.  Similar limitations apply to the principles for collective systems (e.g., design for emergence, homogeneity-heterogeneity).  On the one hand we know that in biological organisms intelligence emerged through evolution from very simple organisms.  On the other hand attempting to 'design' for emergence may appear to be counterintuitive.  Nevertheless, Pfeifer and Bongard are convinced that the design principles will "help us to meet the three main goals of artificial intelligence, i.e., finding general principles of intelligent behavior, building intelligent artifacts, and understanding biological systems" (p. 356).  Furthermore, they point out that "with the synthetic approach there is the exiting possibility that we can exploit intelligence as it could be" (p.359).  Because artificial systems are not constraint by some of the limitations of natural evolution they could evolve novel kinds of brains and bodies.

How the body shapes the way we think goes beyond theoretical discussion and offers a wealth of examples that illustrate the practical implications of the theory.  These examples come from robotics (e.g., the quadruped robot Puppy, the 3D biped robot Denise), animal locomotion (e.g., insect walking) and perception (e.g., motion parallax compensation in insect eyes), machine learning (e.g., soccer playing robots, neural networks), social interactions (e.g., mirror neurons, imitating robots), artificial evolution (e.g., design of fuel pipes, satellite antennas, electronic circuits), evolutionary robotics (e.g., block pushers), modular robots (e.g., HYDRA, M-TRAN), and collective intelligence (e.g., ant navigation, flocking algorithms, robot cooperation). Furthermore, an entire section of the book is dedicated to concrete applications of embodied intelligence.  Here the authors show how the perspective of embodiment can shed new light on topics as diverse as "ubiquitous computing, management, the psychology of human memory, and robotic and artificial intelligence technologies in our everyday lives" (p.246).  For example, Pfeifer and Bongard provide a plausible alternative to the storehouse metaphor of memory by showing how a robot (the artificial mouse) can learn a delayed reward task by only picking up on instantaneous correlations and without remembering the behavior that got rewarded.  They further demonstrate how several of their design principles provide the foundation for their constructive and distributed dynamical model of memory.

Overall, Pfeifer and Bongard have written a very interesting book that explains succinctly and in plain English many concepts from philosophy, psychology, and AI.  Each individual chapter begins with an introduction of the main questions that will be tackled, and ends with a useful summary of the main points.  Several focus boxes allow for easy access to additional background information about key concepts (e.g., mind-body problem, dynamical systems, artificial evolution).  The examples provided to illuminate the (at times very complex) concepts are always relevant and help to make the book accessible to a wide audience.  Pfeifer and Bongard stress throughout the book how important the 'frame of reference problem' is: from ant navigation to robot locomotion they show time and time again that phenomena which, from an observers perspective, appear to require complex computing are in fact based on simple processes that exploit the physical properties of the agents and their environment.  Critics of the embodiment approach might object that Pfeifer and Bongard's focus on the similarities of the challenges that are faced by all biological systems leads them to neglect the differences between the intelligence of other animals and human intelligence.  These critics may be skeptical that embodiment alone can explain why no other species evolved a complex language that not only appears to be extremely useful in the struggle for survival but also allows us to remove ourselves (at least mentally) from the constraints of our physical environment and our evolutionary past.  And, while Pfeifer and Bongard gesture towards an embodied solution of the symbol-grounding problem, much work remains to be done to show that all facets of language and intelligence can be explained via embodiment.  In fairness to the authors: they do not claim that they have solved the mind-body problem and stress repeatedly that further research is needed to confirm their theory.  However, they are confident that the implications of their theory will "profoundly change how we view ourselves and the relationship between mind and body." (p. 370).

© 2007 Christina Behme

Christina Behme, MSc (1986, Biology, University Rostock, Germany), MA (2005, Philosophy, Dalhousie University) is currently a PhD student in the philosophy department at Dalhousie University, Halifax, Nova Scotia, Canada. Her research interests are philosophy of mind and psychology, cognitive science, and philosophy of language.

Categories: Psychology, Philosophical